Vehicle Detection in UAV Aerial Images Based on Improved YOLOv3

Abstract

Vehicle detection in UAV aerial images with complex scenes is a challenging task in intelligent transportation systems, as the sizes of vehicles in the images change with the flight height of UAV. When the UAV is far from the ground, the vehicle object become a small object, which makes it difficult to be detected. This paper presents an improved YOLOv3 model with deeper feature extraction network and four different scale detection layers to detect vehicles in aerial images accurately and robustly. When the high-resolution image of UAV aerial is zoomed to $\mathbf{608}\times\mathbf{608}$ as input, the detection speed of improved YOLOv3 is equivalent to original YOLOv3, and the recall rate and AP are significantly increased by 9%, 11% respectively, while the detection precision reaches 97.09%.